БАКТЕРИЯЛАРДЫҢ МИКРОСКОПИЯЛЫҚ БЕЙНЕЛЕРІН КӨПКЛАССТЫ ЖІКТЕУГЕ АРНАЛҒАН ГИБРИДТІ ТЕРЕҢ ОҚЫТУ МОДЕЛІН ӘЗІРЛЕУ

Abstract

This paper addresses the problem of automatic multiclass classification of microscopic images of bacteria. The study utilizes a dataset comprising 2034 microscopic images representing 33 bacterial taxa. To ensure methodological rigor, data integrity was verified and a strict train/validation/test splitting protocol was applied to prevent information leakage. To characterize image quality and structural properties, several proxy features were computed, including brightness, contrast, Shannon entropy, Laplacian variance, and Sobel gradient energy. Statistical analysis using the Kruskal–Wallis test demonstrated that these features exhibit significant inter-class discriminative power. Classification performance was evaluated using both classical machine learning models and state-of-the-art deep learning architectures. In addition, a hybrid deep learning model based on multiple instance learning was proposed to effectively aggregate local structural patterns within microscopic images. Experimental results indicate that the proposed approach improves classification robustness and accuracy.